Distributed Photovoltaic Power Disaggregation Based on Deep Learning

The penetration of load-side distributed photovoltaics (PV) has been increasing in the face of "dual-carbon" targets. However, PV power generation has brought great challenges to the power balance regulation process of the grid, due to its discontinuity and instability. The first and most important impact is to ensure the observability of PV generation through grid operation and maintenance regulation. Currently, distributed PV power generation data for behind-the-meter access needs to be collected separately through system retrofits, which often requires the addition of separate meters, resulting in additional expensive investments. Most of the behind-the-meter PV generation estimation methods usually use net load data for research, which leads to a low estimation accuracy. In this regard, we propose a behind-the-meter PV power disaggregation method which is based on deep learning to solve this problem. Firstly, the capacity of the user is roughly estimated using the methods of existing studies, and then training samples are generated based on the combination of observable PV examples and historical actual load data of the target user; subsequently, an estimation model consisting of cascade convolutional neural network and bidirectional long short-term memory network is trained by using generated samples. Finally, combined with meteorological information, PV power disaggregation can be classified according to the net load data collected by smart ammeters, and the accuracy of this method is significantly improved by testing on the publicly available Ausgrid dataset.